High-orbit multi-source point target cross-domain fusion algorithm based on modal characteristic distillation

Abstract

The invention discloses a high-orbit multi-source point target cross-domain fusion method based on modal characteristic distillation, and relates to the technical field of modern aerospace monitoring. The method comprises the following steps: extracting invariance features related to target essential attributes from visible light, infrared and radar data through a teacher network, and fusing the invariance features to generate teacher knowledge vectors; and taking the teacher knowledge vector as an optimization target of a student network, and performing iterative optimization by adopting a composite loss function including knowledge distillation loss, physical attribute loss, modal alignment loss and task loss, thereby realizing accurate identification of a point target. Through the feature extraction and knowledge distillation mechanism guided by physical constraints, the problem of feature heterogeneity among multi-modal data is effectively solved, the recognition accuracy and tracking stability of point targets such as space debris in a complex space environment are improved, meanwhile, the model is lightweight and efficient, and the on-orbit real-time processing requirement is met.

Description

Technical Field

[0001] This invention relates to the field of modern aerospace monitoring technology, and more specifically to a cross-domain fusion algorithm for high-orbit multi-source point targets based on modal feature distillation. Background Technology

[0002] In the field of modern aerospace monitoring, high-orbit satellites, with their advantages of wide coverage and stable observation cycles, have become the core platform for missions such as space target monitoring, disaster early warning, and environmental monitoring. With the rapid development of sensor technology, the payload system of high-orbit satellites is becoming increasingly diverse. Multimodal devices such as visible light cameras, infrared detectors, and radar can acquire target information from different physical dimensions: visible light data can capture the texture details and geometric shape of the target, infrared data can reflect the thermal radiation characteristics of the target, and radar data can penetrate clouds and fog to obtain the scattering characteristics and motion parameters of the target. This multi-source heterogeneous observation mode provides a data foundation for improving the monitoring accuracy of point targets (such as spacecraft and space debris), but the cross-domain fusion challenge encountered in its application remains a key bottleneck restricting technological breakthroughs.

[0003] The core requirement for high-orbit point target monitoring lies in achieving accurate target identification, continuous tracking, and attribute inversion. However, due to the fundamental differences in the physical principles of different modal data, their feature spaces exhibit significant heterogeneity: the feature distribution of visible light data depends on illumination conditions and target surface reflectivity, the feature expression of infrared data is related to the target temperature field and emissivity, and radar data is affected by factors such as electromagnetic wave polarization and observation angle. This heterogeneity leads to insurmountable feature gaps between multi-source data, making it difficult for traditional fusion methods to effectively mine complementary information between data. Under complex weather conditions, visible light data may become invalid due to cloud cover, while radar data, although functioning normally, has weak correlations with visible light features. Direct splicing or weighted fusion can lead to feature redundancy or even conflicts, ultimately affecting the robustness of target identification and tracking.

[0004] Feature layer fusion has become a research hotspot due to its balance between information preservation and computational efficiency. Traditional feature layer methods can be divided into two categories: one is based on feature transformation, such as Principal Component Analysis (PCA) and Independent Component Analysis (ICA), which maps multimodal features to a low-dimensional common space through linear transformation, but struggles to capture nonlinear feature relationships; the other is based on feature concatenation, which directly concatenates the feature vectors of each modality and inputs them into a classifier. While simple to implement, this method does not consider feature heterogeneity and is prone to the "curse of dimensionality" and feature redundancy. In recent years, the rise of deep learning technology has provided new avenues for feature fusion. Cross-domain mapping methods based on autoencoders and generative adversarial networks (GANs) have been widely used, achieving feature alignment by learning nonlinear mapping relationships between modalities. CycleGAN achieves style transfer from visible light and infrared images by constructing a bidirectional generative network, but the features it generates often lose essential target attribute information, making it difficult to meet the requirements for accurate identification of high-orbit point targets.

[0005] Therefore, how to extract invariant features related to the essential attributes of the target from heterogeneous observations and break through the bottleneck of cross-domain fusion of high-orbit point targets is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of the above problems, the present invention is proposed to provide a cross-domain fusion algorithm for high-orbit multi-source point targets based on modal feature distillation that overcomes or at least partially solves the above problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: First, the point target is defined as a high-orbit space target observed by multiple source sensors; S1. Acquire multimodal observation data from the high-orbit platform, wherein the multimodal observation data includes at least visible light image data, infrared image data, and radar data; S2. Construct and train a modal feature distillation network, which includes a teacher network and a student network. The teacher network is used to extract invariant features related to the essential attributes of point targets from different modal observation data, and the invariant features corresponding to different modal observation data are weighted and fused to output as a teacher knowledge vector. The teacher knowledge vector serves as the optimization target of the student network and participates in the step of optimizing and training the network parameters of the student network. S3, the student network optimized by S2, is used to perform state identification on the multimodal observation data corresponding to the target point to be identified.

[0008] Preferably, the method further includes signal preprocessing of the multimodal observation data, the signal preprocessing including: Adaptive histogram equalization is performed on the visible light image data to enhance contrast; Atmospheric radiation correction is performed on the infrared image data to eliminate background interference; The radar data is subjected to Kalman filtering to smooth radar cross section (RCS) fluctuations, and Doppler frequency offset is linearly corrected to eliminate the effects of ionospheric dispersion. The processed multimodal data is spatiotemporally aligned so that data from different modalities are aligned to the same baseline in time and space.

[0009] Furthermore, the spatiotemporal alignment includes both temporal alignment and spatial alignment: The time alignment is based on the timestamp of the radar data and uses interpolation to unify the time resolution of visible light and infrared data; the spatial alignment is based on the target's orbital parameters to calculate the spatial projection relationship of each modal data and map the target position in the visible light and infrared images to the radar coordinate system.

[0010] Preferably, in step S2, the step of the teacher network extracting invariant features related to the essential attributes of the point target from observation data of different modalities includes: Convolutional operations are performed on visible light image data to extract grayscale distribution features. A spatial attention mechanism is used to extract grayscale statistical features related to the target material. Invariant features strongly correlated with the material are selected, and a visible light modal knowledge vector is output. Convolutional operations are performed on infrared image data to extract spatial distribution features of radiation intensity, and a long short-term memory network is used to capture the temporal dynamic changes of temperature. At the same time, the extracted radiation features are constrained by a physical loss function based on the Stefan-Boltzmann law to establish a physical relationship between them and the target temperature and emissivity, map out the emissivity-related invariant features, and output an infrared modal knowledge vector. Physical loss function: L phys =|I pred -εσT 4 |; Where I is the radiation intensity; ε is the emissivity; σ is the Boltzmann constant; and T is the temperature. The radar cross section (RCS) sequence and the Doppler frequency offset sequence are fused to capture their temporal correlation. The extracted features are correlated with the target's physical size using a constrained loss function based on radar cross section theory. The target's motion pattern is identified, and a radar mode knowledge vector is output. Constraint loss function: L rcs =|RCS pred- 4πr 2 σ|; Where RCS is the radar cross section value; r is the target's equivalent radius; and σ is the radar cross section.

[0011] Furthermore, the step of outputting the teacher knowledge vector includes: The reliability score of each modality knowledge vector is calculated using an attention mechanism. Based on the reliability score, the visible light, infrared, and radar modality knowledge vectors are weighted and fused to generate the teacher knowledge vector.

[0012] Furthermore, the reliability score is calculated based on the feature variance of each modality knowledge vector and its corresponding physical constraint loss value, wherein the smaller the feature variance and the lower the physical constraint loss value, the higher the reliability score.

[0013] Preferably, in step S2, the student network adopts an encoder-decoder architecture: The teacher knowledge vector is nonlinearly transformed and compressed to reduce its dimensionality and map it to a unified feature representation space. The feature components related to the essential attributes of the point target are extracted from the unified feature representation through a self-attention mechanism. A reconstruction vector matching the dimension of the teacher knowledge vector is reconstructed based on the unified feature representation, and the encoding process is optimized through reconstruction loss. The essential attributes include the target material and motion pattern. The unified feature representation is mapped to the target type probability distribution of the target recognition task and / or the target state vector of the tracking task.

[0014] Preferably, the steps for optimizing the network parameters of the student network include: A distillation loss function is constructed. During the training phase of the student network, the teacher knowledge vector is used as the optimization objective. Based on the output value of the distillation loss function, the parameters of the student network are iteratively optimized through backpropagation until the model converges.

[0015] Furthermore, the distillation loss function is a weighted sum of the following four losses: Knowledge distillation loss is measured using KL divergence to determine the difference between the distribution of student network output features and the distribution of teacher network output knowledge vectors. Physical property loss includes classification cross-entropy loss based on material labels and classification cross-entropy loss based on motion pattern labels; Modality alignment loss employs a triplet loss function to reduce the distance between different modal features of the same target in a unified space, and to increase the distance between the same modal features of different targets. Task loss includes cross-entropy loss for target recognition tasks and / or mean squared error loss for target tracking tasks.

[0016] Based on the same technical concept, the present invention also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the program, it implements the steps of the high-orbit multi-source point target cross-domain fusion method based on modal feature distillation as described above.

[0017] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: 1. By constructing a "teacher-student" network architecture, the teacher network extracts modality-invariant features related to the essential attributes of the target material and motion pattern from multimodal data such as visible light, infrared, and radar. The student network, with the help of the knowledge distillation mechanism, maps the heterogeneous features to the representation space by reducing their dimensionality. This significantly reduces the distribution differences between different modal features and achieves deep alignment and fusion of cross-modal features.

[0018] 2. In complex weather conditions or scenarios where some modal data fails, this method can still dynamically evaluate the reliability of each modal feature through the attention mechanism, and combine physical constraint loss to ensure the physical rationality of feature extraction, thereby maintaining high recognition accuracy and tracking stability even in extreme environments.

[0019] 3. By introducing constraint loss based on physical priors such as Stefan-Boltzmann law and radar scattering theory, the network is guided to focus on features that are strongly related to the essential attributes of the target, suppressing non-essential environmental interference and improving the discriminative and generalization capabilities of the fused features.

[0020] 4. The student network adopts an encoder-decoder structure to compress high-dimensional teacher knowledge vectors into a unified low-dimensional representation. While maintaining high performance, it controls the number of model parameters and inference time, meeting the real-time requirements of on-orbit processing on high-orbit satellite platforms. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a scatter plot of the multimodal feature distribution provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the characteristic gap formation mechanism provided in an embodiment of the present invention; Figure 3 This is an overall process framework diagram provided in the embodiments of the present invention; Figure 4 This is a flowchart of the signal preprocessing provided in an embodiment of the present invention; Figure 5 This is a flowchart of the teacher network feature extraction and fusion process provided in this embodiment of the invention; Figure 6 This is a line graph showing the recognition accuracy of different scenes provided in the embodiments of the present invention; Figure 7 This is a comparison diagram of extreme scene trajectories provided in an embodiment of the present invention; Figure 8 This is a bar chart showing the changes in ablation experimental indicators provided in this embodiment of the invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Combination Figure 2 The schematic diagram shown illustrates the formation mechanism of the feature gap. The features of point targets in different modes exhibit significant distribution gaps due to differences in physical principles, which is the core problem that this invention aims to solve.

[0025] Combination Figure 3 The flowchart shown illustrates the overall fusion framework. This embodiment of the invention discloses a cross-domain fusion algorithm for high-orbit multi-source point targets based on modal feature distillation. Firstly, point targets are defined as high-orbit space targets observed by multiple source sensors. The algorithm includes the following steps: S1. Acquire multimodal observation data from the high-orbit platform. The multimodal observation data includes at least visible light image data, infrared image data, and radar data. S2. Construct and train a modal feature distillation network, which includes a teacher network and a student network. The teacher network is used to extract invariant features related to the essential attributes of the point target from different modal observation data, and the invariant features corresponding to different modal observation data are weighted and fused to output the teacher knowledge vector. The teacher knowledge vector serves as the optimization target of the student network and participates in the step of optimizing and training the network parameters of the student network. S3, the student network optimized by S2, is used to perform state identification on the multimodal observation data corresponding to the target point to be identified.

[0026] In one embodiment, such as Figure 4 As shown, signal preprocessing is performed on the acquired multimodal data, including: Adaptive histogram equalization is performed on visible light image data to enhance contrast; atmospheric radiation correction is performed on infrared image data to eliminate background interference; Kalman filtering is performed on radar data to smooth radar cross section (RCS) fluctuations, and linear correction is performed on Doppler frequency offset to eliminate the effects of ionospheric dispersion. The processed multimodal data is spatiotemporally aligned so that data from different modalities are aligned to the same baseline in time and space.

[0027] Furthermore, spatiotemporal alignment includes both temporal alignment and spatial alignment: Time alignment uses the timestamp of radar data as a reference and interpolates to unify the time resolution of visible light and infrared data; spatial alignment calculates the spatial projection relationship of each modal data based on the target's orbital parameters and maps the target position in the visible light and infrared images to the radar coordinate system.

[0028] To be more specific: Time alignment uses the timestamp of the "Space Fence" radar data as a benchmark, and interpolation is used to unify the time resolution of visible light and infrared data to 1Hz, with synchronization error controlled within ±0.1 seconds; Spatial alignment calculates the spatial projection relationship of each modal data based on the orbital elements of space debris, and maps the target position in the visible light and infrared images to the radar coordinate system, ensuring that the spatial coordinate deviation of the same target in multimodal data is less than 1 pixel. Visible light image data is denoised using the BM3D (Block-Matching3D) algorithm to remove cosmic background noise (signal-to-noise ratio range 8-18dB) while preserving target grayscale features; infrared image data is denoised using wavelet thresholding to eliminate space radiation interference and using moving average filtering to smooth high-frequency fluctuations in the temperature sequence; radar data is denoised using Kalman filtering to estimate the RCS trend term, removes pulse noise caused by space plasma, and eliminates ionospheric dispersion effects through linear correction of Doppler frequency offset. Combination Figure 1 The scatter plot of multimodal feature distribution shown indicates that the features of visible light, infrared and radar modes are scattered and have low overlap in the original space, with significant feature gaps. This further highlights the necessity of cross-domain feature fusion.

[0029] like Figure 5 As shown, in one embodiment, step S2, in which the teacher network extracts invariant features related to the essential attributes of the point target from observation data of different modalities, includes: Convolutional operations are performed on visible light image data to extract grayscale distribution features. A spatial attention mechanism is used to extract grayscale statistical features related to the target material. Invariant features strongly correlated with the material are selected, and a visible light modal knowledge vector is output. Convolutional operations are performed on infrared image data to extract spatial distribution features of radiation intensity, and a long short-term memory network is used to capture the temporal dynamic changes of temperature. At the same time, the extracted radiation features are constrained by a physical loss function based on the Stefan-Boltzmann law to establish a physical relationship between them and the target temperature and emissivity, map out the emissivity-related invariant features, and output an infrared modal knowledge vector. Physical loss function: L phys =|I pred -εσT 4 |; Where I is the radiation intensity; ε is the emissivity; σ is the Boltzmann constant; and T is the temperature. The radar cross section (RCS) sequence and the Doppler frequency offset sequence are fused to capture their temporal correlation. The extracted features are correlated with the target's physical size using a constrained loss function based on radar cross section theory. The target's motion pattern is identified, and a radar mode knowledge vector is output. Constraint loss function: L rcs =|RCS pred -4πr 2 σ|; Where RCS is the radar cross section value; r is the target's equivalent radius; and σ is the radar cross section.

[0030] like Figure 4 As shown, specifically, the teacher network adopts a multi-branch parallel structure, including a visible light branch, an infrared branch, and a radar branch. Each branch extracts features of modal characteristics and outputs modal invariance features. The core of the visible light branch is to extract material-related invariant features from light spot-like targets. Its structure includes a convolutional feature extraction module, a material-aware attention module, and a feature selection layer. The convolutional feature extraction module consists of three convolutional layers with kernel sizes of 3×3, 3×3, and 1×1. It extracts grayscale distribution features through a sliding window with a stride of 1. The first convolutional layer learns local grayscale differences, the second convolutional layer captures grayscale gradient changes, and the third 1×1 convolutional layer compresses the channel dimension (from 64 dimensions to 32 dimensions) to reduce redundant information. The material-aware attention module introduces material-grayscale mapping priors (e.g., high reflectivity of metallic materials corresponds to high grayscale mean). It strengthens material-related grayscale statistical features (mean, entropy) through a spatial attention mechanism to suppress grayscale fluctuations caused by illumination changes. The attention weight map output by this module can visualize the model's focus area on material features. The feature selection layer maps 32-dimensional features to 16-dimensional features through a fully connected layer and uses L1 regularization to constrain sparsity, selecting invariant features strongly related to materials (e.g., grayscale mean, gradient variance), and outputs a visible light modality knowledge vector. The infrared branch focuses on extracting invariant features related to temperature and emissivity. Its structural design includes a radiation feature encoding module, a thermodynamic constraint module, and an emissivity sensing module. The radiation feature encoding module consists of two convolutional layers and one long short-term memory (LSTM) layer. The convolutional layers extract the spatial distribution features of infrared radiation intensity, and the LSTM layer captures the dynamic changes in temperature over time (such as periodic temperature fluctuations in spacecraft), outputting a 64-dimensional temporal feature vector. The thermodynamic constraint module embeds the Stefan-Boltzmann law I=εσT. 4 Where I is the radiation intensity, ε is the emissivity, σ is the Boltzmann constant, and T is the temperature, through the physical loss function L phys =|I pred -εσT 4 |Constrained feature learning ensures the physical correlation between the extracted radiation features and temperature and emissivity; the emissivity sensing module maps 64-dimensional temporal features to the emissivity-related feature subspace through a fully connected layer, and outputs a 16-dimensional infrared modal knowledge vector, since emissivity is an invariant parameter related to material (e.g., metal ε≈0.1, composite material ε≈0.8).

[0031] The radar branch focuses on extracting invariant features related to target size, shape, and motion. Its structure includes an RCS-Doppler feature fusion module, an electromagnetic scattering constraint module, and a motion pattern classification module. The RCS-Doppler feature fusion module processes the RCS sequence (1D) and the Doppler frequency offset sequence (1D) through a dual-channel convolutional network with a kernel size of 5×1, capturing their temporal correlation (e.g., the increase in Doppler frequency offset accompanied by RCS fluctuations during acceleration), outputting a 32-dimensional fused feature. The electromagnetic scattering constraint module introduces the radar cross section theory RCS=4πr²σ, where r is the target's equivalent radius and σ is the radar cross section, constrained by the loss function L. rcs =|RCS pred -4πr²σ| ensures the correlation between features and target physical size, reducing clutter interference; the motion mode classification module uses a 3-layer fully connected network to classify 32-dimensional fused features, identify the target's motion mode (uniform speed, acceleration, deceleration), and use the classification probability distribution as part of the knowledge vector, finally outputting a 16-dimensional radar mode knowledge vector.

[0032] In one embodiment, the reliability score of each modality knowledge vector is calculated through an attention mechanism, and the visible light, infrared, and radar modality knowledge vectors are weighted and fused based on the reliability scores to generate the teacher knowledge vector.

[0033] Furthermore, the reliability score is calculated based on the feature variance of each modality's knowledge vector and its corresponding physical constraint loss value. The smaller the feature variance and the lower the physical constraint loss value, the higher the reliability score.

[0034] Specifically: The 16-dimensional knowledge vectors (out of a total of 48 dimensions) output from the three branches are dynamically adjusted using an attention mechanism; a reliability score (based on feature variance and physical constraint loss value) is calculated for each modality knowledge vector; the smaller the variance and the lower the loss value, the higher the reliability; the reliability score is then converted into a weight coefficient ω using a softmax function. vis ,ω ir ,ω radar , satisfying ω vis +ω ir +ω radar =1 weighted fusion of multimodal knowledge vectors, outputting a 48-dimensional teacher network knowledge vector, which serves as the learning objective of the student network.

[0035] In one embodiment, the student network mentioned in step S2 adopts an encoder-decoder architecture: The teacher's knowledge vector is nonlinearly transformed and compressed to reduce its dimensionality and map it to a unified feature representation space. The feature components related to the essential attributes of the point target are extracted from this unified feature representation through a self-attention mechanism. A reconstruction vector matching the dimension of the teacher's knowledge vector is reconstructed based on the unified feature representation, and the encoding process is optimized through reconstruction loss. The essential attributes include the target material and motion pattern. Map the unified feature representation to the target type probability distribution of the target recognition task and / or the target state vector of the tracking task.

[0036] Specifically, the core task of the student network is to learn the invariant features output by the teacher network and form a unified feature representation. Its structure adopts an encoder-decoder architecture. The feature encoding module receives the 48-dimensional knowledge vector output by the teacher network and maps it to a low-dimensional feature space through nonlinear transformation. The compression layer is a two-layer fully connected network that compresses the 48-dimensional vector to 32 dimensions and uses the ReLU activation function to enhance the nonlinear expressive power. The essential attribute focusing module introduces the prior distribution of the target's essential attributes (material, motion mode) and strengthens the features related to material (such as metal / composite materials) and motion mode (such as uniform speed / acceleration) through a self-attention mechanism to suppress mode-specific noise. The bottleneck layer is a one-layer fully connected network that compresses the 32-dimensional features to 16 dimensions, which serves as the core of the unified feature representation. This dimension is chosen to balance feature expressive power and computational efficiency.

[0037] The feature decoding module is used to verify the effectiveness of the encoded features. It evaluates the degree of feature retention by reconstructing the teacher network knowledge vector: the expansion layer is a 2-layer fully connected network that expands the 16-dimensional core features to 48 dimensions, consistent with the dimensions of the teacher network knowledge vector; the reconstruction loss layer calculates the mean squared error (MSE) between the reconstructed vector and the teacher knowledge vector, and optimizes the encoding module through backpropagation to ensure that the core features contain sufficient invariant features.

[0038] The output layer maps the 16-dimensional core features to the feature space required for target recognition and tracking: for recognition tasks, the softmax layer outputs the probability distribution of target types (spacecraft, space debris, missiles); for tracking tasks, the output target state vector (position, velocity, acceleration) is used for subsequent trajectory prediction.

[0039] In one embodiment, the steps of optimizing the network parameters of the student network include: A distillation loss function is constructed. During the training phase of the student network, the teacher's knowledge vector is used as the optimization objective. Based on the output value of the distillation loss function, the parameters of the student network are iteratively optimized through backpropagation until the model converges.

[0040] Furthermore, the distillation loss function is a weighted sum of the following four losses: Knowledge distillation loss is measured using KL divergence to determine the difference between the distribution of student network output features and the distribution of teacher network output knowledge vectors. Physical property loss includes classification cross-entropy loss based on material labels and classification cross-entropy loss based on motion pattern labels; Modality alignment loss employs a triplet loss function to reduce the distance between different modal features of the same target in a unified space, and to increase the distance between the same modal features of different targets. Task loss includes cross-entropy loss for target recognition tasks and / or mean squared error loss for target tracking tasks.

[0041] Specifically, the distillation loss function is key to students' online learning of teacher knowledge. It needs to simultaneously constrain feature distribution alignment, physical attribute consistency, and task performance, and consists of four parts: knowledge distillation loss, physical attribute loss, modality alignment loss, and task loss.

[0042] Knowledge distillation loss L kd To ensure that the feature distribution of the student network output is consistent with that of the teacher network, the KL divergence is used to measure the difference in their probability distributions: L kd =KL(P teacher ||P student ), where P teache It is the probability distribution of the teacher network knowledge vector (normalized by softmax), P student It is the probability distribution of the student network's coding features, and this loss forces the student network to learn the cross-modal association patterns captured by the teacher network.

[0043] Physical property loss L phyConstraining the consistency between student network features and target essential attributes includes material loss and motion loss: Material loss is based on material labels (metal / composite materials) inferred from visible light and infrared features, calculating classification cross-entropy loss; Motion loss is based on motion pattern labels (uniform speed / acceleration / deceleration) inferred from radar and infrared features, calculating classification cross-entropy loss; its expression is L phy =α×L material +β×L motion , where α and β are weighting coefficients (α=0.6, β=0.4), which are adjusted according to the importance of the attribute.

[0044] Modal alignment loss L align To reduce the distance differences between multimodal features in a unified space, triplet loss is employed: L align =max(d pos -d neg +margin,0), where d pos It is the distance between different modal features of the same target, d neg It is the distance between the same modal features of different targets, and the margin is the threshold (set to 1.0). This loss ensures that the multimodal features of the same target are clustered in a unified space, and the features of different targets are separated.

[0045] Mission loss L task To directly optimize target recognition and tracking performance, cross-entropy loss (Li) is used for recognition tasks. cls ), measures the difference between the predicted category and the true label; for tracking tasks, mean squared error loss (L) is used. reg The measure of the difference between the predicted state and the actual trajectory is L; its expression is L. task =γ×L cls +δ×L reg γ and δ are dynamic weights that are adjusted according to task priority (γ=0.7, δ=0.3 for identification tasks; γ=0.3, δ=0.7 for tracking tasks).

[0046] The total loss function is the weighted sum of the losses of each component: L total =λ1×L kd +λ2×L phy +λ3×L align +λ4×L task , where λ1=0.3, λ2=0.2, λ3=0.2, λ4=0.3.

[0047] The results were optimized through validation set experiments to ensure the synergistic effect of each loss term.

[0048] It should be noted that the distillation loss function is explicitly used in the model training mechanism of the student network: by continuously modifying the control parameters of the student network, the student network is forced to imitate the teacher network at the internal feature level, thereby effectively transferring and distilling the deep, target-related invariant knowledge learned by the teacher network from multimodal observation data into the student network.

[0049] Based on the same technical concept, the present invention also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the program to implement the steps of any of the above-described methods for cross-domain fusion of high-orbit multi-source point targets based on modal feature distillation.

[0050] Furthermore, to verify the effectiveness of the high-orbit multi-source point target cross-domain fusion framework based on modal feature distillation, the following describes the construction method of the experimental dataset, including data sources, preprocessing procedures and scene division, and clarifies the various evaluation indicators used to assess model performance, providing a standardized verification benchmark for subsequent comparative experiments. Meanwhile, the construction of multi-source target datasets faces the dual challenges of data scarcity and scenario complexity. Due to the limitations of satellite mission planning and security requirements, real space debris monitoring data is difficult to acquire on a large scale. Therefore, this study adopts a hybrid construction mode of "real data + simulation data" to ensure the physical authenticity of the data while covering diverse and complex scenarios. Visible light data: Sourced from the visible light camera payload of a U.S. space surveillance satellite, primarily monitoring space debris in geosynchronous orbit (GEO) and medium Earth orbit (MEO) regions, with a resolution of 30 meters. It includes 3,000 frames of point target images from January to June 2023, with target types including abandoned satellites, rocket debris, and other debris. The image size is 512×512 pixels.

[0051] Infrared data: The data is from the infrared detectors of the European Space Agency's Space Situation Awareness Program satellite, with a wavelength range of 8-14μm (long-wave infrared) and a time resolution of 30 minutes. It includes 2,000 sets of space debris radiation intensity sequences and simultaneously records the target's orbital parameters and temperature markings. Radar data: Taken from actual measurements of the U.S. Space Fence radar system, operating at the S-band (2-4 GHz), mainly for monitoring space debris in low Earth orbit (LEO) and high orbit regions. It includes RCS sequences (sampling rate 1 Hz) and Doppler frequency offset data of 1,500 point targets, with the equivalent radius of the targets ranging from 0.1 to 5 meters. A multimodal simulation platform for space debris was built based on STK and MATLAB. Simulation data was generated according to the target material (metal, composite material, non-metal), motion mode (uniform speed, rotation, tumbling) and space environment parameters (solar radiation intensity, space plasma density, micrometeoroid impact probability). Visible light simulation: This project simulates the reflection characteristics of sunlight by space debris using a ray tracing algorithm, introduces cosmic background noise to simulate the impact of the space environment, and generates 5000 frames of point target images under different lighting conditions. Infrared simulation: Based on the thermodynamic model, the temperature field distribution of space debris in a vacuum environment is calculated. Combined with the solar radiation heating model, the temperature change is simulated, generating 4,000 sets of radiation intensity sequences, covering a temperature range from -200℃ to 100℃. Radar simulation: The physical optics (PO) method is used to calculate the RCS of space debris. A space plasma interference model is introduced to simulate radar echo distortion, generating 3000 sets of RCS and Doppler data, including characteristic changes under different attitude angles (0°-360°) and rotation states. Table 1. Detailed Distribution of Dataset Samples

[0052] To ensure the spatiotemporal consistency and feature validity of multimodal data, the preprocessing procedure includes the following steps: Time alignment: Based on the timestamp of the "Space Fence" radar data, the time resolution of visible light and infrared data is unified to 1Hz through interpolation, and the synchronization error is controlled within ±0.1 seconds; Spatial alignment: Based on the orbital elements of space debris, the spatial projection relationship of each modal data is calculated, and the target position in the visible light and infrared images is mapped to the radar coordinate system to ensure that the spatial coordinate deviation of the same target in multimodal data is less than 1 pixel; Visible light data: The BM3D (Block-Matching3D) denoising algorithm is used to remove cosmic background noise (signal-to-noise ratio range of 8-18dB) while preserving the grayscale features of the target. Infrared data: spatial radiation interference is eliminated by wavelet threshold denoising, and high-frequency fluctuations in the temperature sequence are smoothed by moving average filtering; Radar data: The RCS trend term is estimated based on Kalman filtering, and the pulse noise caused by space plasma is removed. The Doppler frequency offset is eliminated by linear correction to eliminate the influence of ionospheric dispersion. Target attribute labeling: The material (metal / composite material / non-metal), motion mode (uniform speed / rotation / tumbling), and target type (destroyed satellite / rocket debris / fragmentation) of the target are labeled by experts in the field of space debris, and the labeling accuracy rate must reach more than 95%; Track annotation: Based on precise orbital data, the position (right ascension, declination, distance), velocity, and acceleration of space debris at each moment are annotated as ground truth for the tracking task. Furthermore, to comprehensively evaluate the robustness of the model in complex spatial environments, the dataset was divided into three categories according to scene complexity, and the sample distribution for each category is shown in Table 2: Typical scenario: No significant space environment interference, stable illumination (sun angle with target 30°-60°), low space plasma density (<10). 8 m -3 ) contains 30% of the sample size and is used to evaluate the basic performance of the model. Interference scenarios: Single spatial environmental interference exists, such as visible light data being affected by solar flares (sudden changes in illumination), infrared data being affected by deep space background radiation (signal-to-noise ratio <12dB), and radar data being affected by plasma sheath interference. This includes 50% of the sample size and is used to evaluate the model's resistance to specific interferences. Extreme scenarios: Multiple sources of interference superimposed, such as a strong solar storm (solar wind speed > 800 km / s) + high-speed tumbling of a target (angular velocity > 5° / s) + high-density space debris (density > 10). -6 km -3 ), which includes 20% of the sample size, is used to test the model's extreme performance. Table 2 Dataset Scene Division and Sample Distribution

[0053] Furthermore, for the task of multi-source target identification and tracking in space debris surveillance, evaluation indicators are designed from four dimensions: feature fusion effect, identification performance, tracking accuracy, and computational efficiency, as follows: First, the feature fusion performance index is quantitatively characterized, including the following calculation methods: Cross-Modal Similarity (CMS) is defined as a measure of the similarity between features from different modalities in a unified representation space, expressed as the mean of cosine similarity.

[0054] in, , , denoted as visible light, infrared, and radar feature vectors for the i-th target, respectively, and N is the number of samples. A higher CMS value (range [-1, 1]) indicates better consistency in the fusion of multimodal features and a more significant mitigation effect on the "feature gap".

[0055] Invariant Feature Retention (IFR) is defined as the proportion of features in the fused features that are related to the essential attributes of the target (material, motion pattern), and is measured by mutual information.

[0056] Among them, the higher the IFR value (range [0,1]), the richer the invariant knowledge preserved by the fused features, and the more accurate the characterization of the essential attributes of the target.

[0057] Secondly, the target recognition performance indicators are quantitatively characterized, including the following calculation methods: Accuracy (Acc) is defined as the proportion of correctly identified target samples out of the total number of samples. For multi-classification tasks involving target types (destroyed satellites / rocket debris / fragments):

[0058] in, TP It is a true positive (a correctly identified positive sample). TN It is a true negative (a correctly identified negative sample). FP It was a false positive. FN This is a false negative. The data result... ACC It directly reflects the model's ability to distinguish target types; the higher the better.

[0059] A confusion matrix is ​​defined as a matrix representation of the recognition results for each category of target. Rows represent the true categories, columns represent the predicted categories, diagonal elements represent the number of correctly identified targets, and off-diagonal elements represent the number of confused targets. It provides a visual analysis of the model's bias in recognizing specific categories of targets, such as whether it is prone to misidentifying small debris as rocket wreckage.

[0060] The F1 score, defined as the harmonic mean of precision and recall, is used to balance imbalanced positive and negative samples.

[0061] Its significance lies in comprehensively evaluating the robustness of the model's recognition capabilities, making it particularly suitable for small-sample target recognition in extreme scenarios.

[0062] The target tracking performance metrics are further quantified, including the following calculation methods: The root mean square error (RMSE) is defined as the root mean square deviation between the predicted and actual positions, and is calculated in the celestial coordinate system.

[0063] in, The true position (right ascension, declination, distance) at time t. To predict the location, T This refers to the tracking duration. The significance of this calculation lies in its role as a core indicator for measuring tracking accuracy; distance error is measured in meters, and angle error in arcseconds. The smaller the value, the better. Trajectory smoothness (TS) is defined as the degree of smoothness of a trajectory, assessed by the variance of the predicted angular velocity.

[0064] in, for Predicted angular velocity at time t. This represents the average angular velocity. The significance of this calculation lies in... The smaller the value, the less noise interference the tracking trajectory is affected, and the more stable the motion state estimation is.

[0065] Tracking Duration Rate (TDR) is defined as the proportion of successful tracking time to total tracking time. Tracking is considered to have failed when the position error exceeds a threshold (50 meters).

[0066] The significance of this calculation is to assess the tracking robustness of the model in complex spatial scenarios, especially suitable for scenarios where the target is tumbling at high speed or the signal is interrupted.

[0067] Finally, the computational efficiency index is quantitatively characterized, including the following calculation methods: Average Inference Time (AIT) is defined as the average time required for a model to process a single set of multimodal data (visible light + infrared + radar), measured in milliseconds. Its purpose is to measure the model's real-time performance, which is crucial for on-orbit processing tasks of space debris monitoring satellites and must be kept below 100ms. Model parameters (MP) are defined as the total number of trainable parameters in a model, expressed in millions (M). The purpose is to reflect the lightweight nature of the model; the smaller the number of parameters, the easier it is to deploy on resource-constrained space surveillance satellite platforms.

[0068] To comprehensively evaluate the performance of this invention, the following four methods were selected as comparative benchmarks: Traditional feature concatenation (FC) method: directly concatenates visible light, infrared, and radar feature vectors and then inputs them into the classifier and tracker; Principal Component Analysis-Based Feature Fusion (PCA-Fusion): Multimodal features are reduced in dimensionality and linearly transformed using PCA, then mapped to a common feature space and fused. Cross-domain mapping based on generative adversarial networks (GAN-Fusion): CycleGAN is used to construct bidirectional generative networks for visible light-infrared and infrared-radar, realizing feature transfer and fusion between modes; Attention-based modal weighted fusion: The dynamic weights of each modality feature are calculated through a self-attention mechanism, and the fused feature is output after weighted summation. This invention is denoted as Modal Feature Distillation Fusion (MFD-Fusion). All methods were implemented under the same experimental environment to ensure fairness in the comparison. The experimental hardware environment was: Intel Xeon Gold 6248 processor (2.5GHz), NVIDIA Tesla V100 graphics card (32GB VRAM), 64GB RAM; the software environment was Python 3.8 and PyTorch 1.10 deep learning framework. Model training parameter settings: batch size of 32, 100 training epochs, initial learning rate of 0.001, using the Adam optimizer (β1=0.9, β2=0.999), with the learning rate decaying to 0.5 every 20 epochs. The weight coefficients of the distillation loss function were set according to Chapter 4 (λ1=0.3, λ2=0.2, λ3=0.2, λ4=0.3). All methods were trained on the training set, hyperparameters were tuned on the validation set, and performance was evaluated on the test set. The experiments were conducted in the three scenarios (normal scenarios, interference scenarios, and extreme scenarios) as defined above. The dataset was divided into a training set (7700 sets), a validation set (1100 sets), and a test set (2200 sets) in a ratio of 7:1:2 to ensure that the distribution of each scenario in the test set was consistent with that in the original dataset. As shown in Table 3, it illustrates the performance of different methods in cross-modal similarity (CMS) and invariant feature retention rate (IFR).

[0069] Table 3 Comparison of Feature Fusion Performance Indicators

[0070] The results show that the proposed MFD-Fusion method significantly outperforms other methods in terms of CMS and IFR. The CMS reaches 0.76, which is 33.3% higher than Attention-Fusion, indicating that it effectively alleviates the "feature gap" and achieves deep alignment of multimodal features. The IFR reaches 0.83, indicating that more invariant knowledge related to target material and motion pattern is retained in the fused features, verifying the effectiveness of the "modal feature distillation" mechanism.

[0071] Traditional FC methods have the lowest CMS and IFR because they do not handle feature heterogeneity and only perform simple concatenation. Although PCA-Fusion improves feature consistency through linear transformation, it cannot capture non-linear associations. Although the features generated by GAN-Fusion are aligned in modal style, they are prone to losing essential attribute information, resulting in an IFR lower than MFD-Fusion. Attention-Fusion improves the fusion effect through dynamic weights, but it does not fundamentally solve the feature heterogeneity problem, and its performance is still inferior to the proposed methods. like Figure 6 As shown in the figure, the recognition accuracy Acc of each method is displayed in different scenarios, where: Typical scenarios: FC (72.3%), PCA-Fusion (78.5%), GAN-Fusion (82.1%), Attention-Fusion (85.7%), MFD-Fusion (93.2%) Interference scenarios: FC (61.5%), PCA-Fusion (67.8%), GAN-Fusion (73.2%), Attention-Fusion (78.9%), MFD-Fusion (89.4%) Extreme scenarios: FC (48.2%), PCA-Fusion (53.7%), GAN-Fusion (59.3%), Attention-Fusion (65.8%), MFD-Fusion (82.6%) Furthermore, Table 4 presents the F1 score and confusion matrix analysis under extreme scenarios.

[0072] Table 4 Recognition performance indicators under extreme scenarios

[0073] Analysis shows that MFD-Fusion has the best recognition performance in all scenarios, especially in extreme scenarios: the accuracy reaches 82.6%, which is 25.5% higher than Attention-Fusion, and the F1 score reaches 0.80, indicating that it can still maintain robust recognition ability in scenarios with small samples and high interference. From the perspective of scene adaptability, the accuracy of all methods decreases with increasing environmental complexity, but MFD-Fusion shows the smallest decrease (10.6% decrease from normal to extreme scenarios), while the FC method decreases by 24.1%, verifying the strong adaptability of the proposed method to complex space environments. Confusion matrix analysis shows that MFD-Fusion has the strongest ability to distinguish between small debris and rocket debris (error rate of only 4.3%), because it retains material and structure-related invariant characteristics through a distillation mechanism, reducing interference from apparent features.

[0074] Table 5. Balance of computational efficiency and performance for each method

[0075] Table 5 shows the tracking accuracy and robustness metrics of different methods. The average inference time in the table represents the time spent processing a single set of multimodal data, the number of model parameters reflects network complexity, and Acc in extreme scenarios represents the target recognition accuracy under high interference conditions; higher values ​​indicate better performance. Data comparison shows that while MFD-Fusion is slightly higher than FC, PCA-Fusion, and Attention-Fusion in terms of average inference time and number of model parameters, it combines… Figure 7 As shown, the present invention has a significantly higher recognition accuracy in extreme scenarios, achieving a better balance between efficiency and performance, and is better able to meet the engineering application needs of high-orbit multi-source target monitoring.

[0076] Table 6 Comparison of Target Tracking Performance Indicators

[0077] As shown in Table 6, MFD-Fusion's predicted trajectory has the highest overlap with the actual trajectory and small position error fluctuations. Attention-Fusion's trajectory shows obvious drift, especially during the high-speed tumbling phase of the target. GAN-Fusion's trajectory has good smoothness, but the cumulative error is large. FC and PCA-Fusion's trajectories are messy and tracking is easily interrupted. The results show that the position RMSE of MFD-Fusion is only 32.5 meters, which is 44.2% lower than that of Attention-Fusion. It also has the best trajectory smoothness (TS=0.05) and tracking persistence rate (TDR=92.7%), indicating that it can stably output high-precision trajectory and maintain tracking continuity even under the interference of high-speed target tumbling and strong solar storm. Traditional methods suffer from large tracking errors and are prone to losing targets due to feature heterogeneity. Although GAN-Fusion improves feature consistency through style transfer, the physical accuracy of the generated features is insufficient, resulting in large cumulative errors. MFD-Fusion, on the other hand, significantly improves tracking robustness by distilling invariant features related to motion patterns (such as angular velocity and acceleration patterns). Table 7 Comparison of computational efficiency indicators

[0078] Table 7 shows the average inference time and model parameter count for each method. The results indicate that FC and PCA-Fusion have the highest computational efficiency but the worst performance. GAN-Fusion has the longest inference time (142.3ms) due to its complex generator network structure, making it difficult to meet real-time requirements. MFD-Fusion has an average inference time of 45.8ms, which is higher than FC and Attention-Fusion, but still within the real-time threshold (100ms) for satellite on-orbit processing, and its parameter count (18.5M) is moderate, making it easy to deploy. Considering both performance and efficiency, MFD-Fusion achieves a good balance. Furthermore, to clarify the role of each module in the framework, this invention designed an ablation experiment, gradually removing three key modules: the teacher network physical constraints, the student network attention mechanism, and the distillation loss function, and comparing their impact on fusion performance. The results are shown in Table 8, including the accuracy indicators. Figure 8 The bar chart is shown (using Acc and RMSE as evaluation metrics under extreme scenarios). Table 8 Ablation Experiment Results

[0079] After removing physical constraints, the recognition accuracy decreased by 7.3%, while the location RMSE increased by 8.7 meters. Analysis shows that physical constraints (such as the Stefan-Boltzmann law and radar scattering theory) can guide the teacher network to focus on features that conform to physical laws, reducing interference from "spurious correlation" features. For example, in infrared feature processing, physical constraints ensure the consistency of radiation intensity and temperature, preventing the model from misclassifying noise as target features, thereby improving feature reliability. Removing the attention mechanism resulted in a more significant performance degradation (accuracy -10.5%, RMSE +13.3 meters). The attention mechanism enables the student network to automatically focus on cross-modal common knowledge (such as the consistent characteristics of metal materials across modalities) and suppress modality-specific noise (such as visible light illumination fluctuations). Without this mechanism, the student network treats all features equally, leading to an increase in redundant information in the fused features and a performance decrease.

[0080] Removing the distillation loss function resulted in a significant performance drop (accuracy -19.1%, RMSE +34.8m), indicating that it is a core module of the framework. The distillation loss achieves deep alignment of cross-modal features by constraining the student network to learn invariant knowledge from the teacher network. If only traditional supervised losses (such as classification losses) are used, the student network struggles to overcome the limitations of modal differences, and the fusion effect is close to Attention-Fusion, validating the necessity of the "teacher-student" distillation architecture. Comparative experiments show that the proposed MFD-Fusion method significantly outperforms traditional methods and mainstream cross-domain fusion algorithms in feature fusion, target recognition, and tracking performance. It exhibits strong robustness, especially in extreme scenarios, validating the effectiveness of the "modal feature distillation" mechanism in mitigating the "feature gap." Ablation experiments further reveal that the distillation loss function is the core of the framework, responsible for the transfer of cross-modal knowledge; the student network attention mechanism enhances the focus on invariant features; and the teacher network physical constraints ensure the physical reliability of the features. The synergistic effect of these three elements achieves high-performance fusion. In summary, the experimental results fully demonstrate that by extracting invariant features related to the essential attributes of the target, the proposed method breaks through the limitations of traditional feature splicing and provides an efficient solution for cross-domain fusion of high-orbit multi-source targets.

[0081] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0082] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A high orbit multi-source point target cross-domain fusion method based on modal feature distillation, characterized in that, The point target is a high-orbit space target observed by a multi-source sensor; comprising the following steps: S1, obtaining multi-modal observation data from a high-orbit platform, the multi-modal observation data at least including visible light image data, infrared image data and radar data; S2, constructing and training a modal feature distillation network, the modal feature distillation network including a teacher network and a student network; the teacher network is used to extract invariance features related to the essential properties of the point target from different modal observation data respectively, and to output a teacher knowledge vector after weighted fusion of the invariance features corresponding to different modal observation data; the teacher knowledge vector is used as an optimization target of the student network, and is used in the step of network parameter optimization training of the student network; S3, the student network optimized through S2 is used for state recognition of the multi-modal observation data corresponding to the point target to be identified.

2. The high-orbit multi-source point target cross-domain fusion method based on modal feature distillation according to claim 1, characterized in that, Further comprising signal preprocessing of the multi-modal observation data, the signal preprocessing including: performing adaptive histogram equalization on the visible light image data to enhance contrast; performing atmospheric radiation correction on the infrared image data to eliminate background interference; performing Kalman filtering on the radar data to smooth the radar scattering cross section (RCS) fluctuation, and performing linear correction on the Doppler frequency offset to eliminate the ionospheric dispersion effect; aligning the processed multi-modal data in time and space to align the data of different modalities in time and space to the same reference.

3. The high-orbit multi-source point target cross-domain fusion method based on modal feature distillation according to claim 2, characterized in that, The time and space alignment includes time alignment and space alignment: The time alignment is to unify the time resolution of visible light and infrared data through interpolation method based on the time stamp of radar data; The space alignment is to calculate the space projection relationship of each modal data based on the orbit parameters of the target, and map the target position in the visible light and infrared image to the radar coordinate system.

4. The high-orbit multi-source point target cross-domain fusion method based on modal feature distillation according to claim 1, characterized in that, In step S2, the step of extracting invariance features related to the essential properties of the point target from different modal observation data by the teacher network includes: performing convolution operation on the visible light image data to extract gray distribution features, extracting gray statistical features related to target material through spatial attention mechanism, screening out invariance features strongly related to material, and outputting visible light modal knowledge vector; performing convolution operation on the infrared image data to extract spatial distribution features of radiation intensity, capturing the time sequence dynamic change of temperature through long short-term memory network; at the same time, using a physical loss function based on Stefan-Boltzmann law to constrain the extracted radiation features, establishing the physical correlation between the extracted radiation features and target temperature and emissivity, mapping the invariance features related to emissivity, and outputting infrared modal knowledge vector; Physical loss function: L phys = |I pred - εσT 4 |; where I is the radiation intensity, ε is the emissivity, σ is the Boltzmann constant, and T is the temperature; fusing the radar scattering cross section (RCS) sequence and the Doppler frequency offset sequence to capture the time correlation of the two; using a constraint loss function based on radar scattering cross section theory to establish the correlation between the extracted features and the physical size of the target, identifying the motion mode of the target, and outputting the radar modal knowledge vector; Constrained loss function: L rcs = |RCS pred - 4πr 2 σ|; where RCS is the radar scattering cross section value, r is the equivalent radius of the target, and σ is the radar cross section.

5. The high-orbit multi-source point target cross-domain fusion method based on modal feature distillation according to claim 4, characterized in that, The step of outputting the teacher knowledge vector comprises: The reliability score of each modality knowledge vector is calculated through an attention mechanism, and the teacher knowledge vector is generated by weighted fusion of the visible light, infrared, and radar modality knowledge vectors according to the reliability score.

6. The high-orbit multi-source point target cross-domain fusion method based on modal feature distillation according to claim 5, characterized in that, The calculation of the reliability score is based on the feature variance of each modality knowledge vector and the corresponding physical constraint loss value, wherein the smaller the feature variance and the lower the physical constraint loss value, the higher the reliability score.

7. The method of claim 1, wherein, In step S2, the student network adopts an encoder-decoder architecture: The teacher knowledge vector is nonlinearly transformed and compressed, and is mapped to a unified feature representation space, and the feature components related to the essential attributes of the point target are extracted from the unified feature representation through a self-attention mechanism; A reconstructed vector matching the dimension of the teacher knowledge vector is reconstructed according to the unified feature representation, and the encoding process is optimized through a reconstruction loss; the essential attributes include target material and motion mode; The unified feature representation is mapped to a target type probability distribution of a target recognition task and / or a target state vector of a tracking task.

8. The high-orbit multi-source point target cross-domain fusion method based on modal feature distillation according to claim 1, characterized in that, The step of optimizing and training the network parameters of the student network comprises: A distillation loss function is constructed, and in the training stage of the student network, the teacher knowledge vector is taken as an optimization target, and based on the output value of the distillation loss function, the parameters of the student network are iteratively optimized through back propagation until the model converges.

9. The high-orbit multi-source point target cross-domain fusion method based on modal feature distillation according to claim 8, characterized in that, The distillation loss function is the weighted sum of the following four losses: The knowledge distillation loss adopts KL divergence to measure the difference between the output feature distribution of the student network and the output knowledge vector distribution of the teacher network; The physical attribute loss includes a classification cross-entropy loss based on a material label and a classification cross-entropy loss based on a motion mode label; The modality alignment loss adopts a triplet loss function to reduce the distance of the same target different modality features in the unified space and increase the distance of different target same modality features; The task loss includes a cross-entropy loss for a target recognition task and / or a mean square error loss for a target tracking task.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the high-altitude multi-source point target cross-domain fusion method based on modality feature distillation according to any one of claims 1-9 when executing the program. The processor implements the steps of the high-altitude multi-source point target cross-domain fusion method based on modality feature distillation according to any one of claims 1-9 when executing the program.

Images